基于电子舌和电子鼻结合CNN-Transformer模型的绿茶种类识别

Green tea species recognition based on electronic tongue and electronic nose combined with CNN-Transformer model

[目的]实现绿茶种类的快速识别。[方法]提出一种基于电子舌和电子鼻结合CNN-Transformer组合模型实现绿茶种类辨识的快速检测方法。分别使用电子舌、电子鼻对5种不同种类绿茶采集味觉、嗅觉的指纹信息,利用短时傅里叶变换(short-time fourier transform,STFT)将一维电子舌和电子鼻信号转换为二维时频图,充分揭示信号能量在时频域的分布特性;提出一种CNN-Transformer组合模型实现电子舌和电子鼻的信息融合和模式识别。该模型引入选择性核卷积和归一化注意力设计卷积模块来替换传统的CNN卷积层,以实现对信号时频图的局部特征动态提取;采用Transformer编码器中的多头自注意力机制提取电子舌和电子鼻特征的全局时序信息,并实现其特征的加权融合;最后,通过全连接层进行分类识别。[结果]基于电子舌和电子鼻的信息融合方法能够有效提取绿茶样本的味觉和嗅觉信号深层特征,并为模型提供更丰富的融合特征表征,以实现对不同绿茶种类的高准确识别,其测试集准确率、精确率、召回率和F 1-Score分别达到99.00%,99.05%,99.00%,99.00%。[结论]试验方法具有成本低、快速、高效等特点。

[Objective]To realize rapid detection of green tea species identification.[Methods]A rapid detection method based on the combination of electronic tongue and electronic nose combined with CNN-Transformer composite model was proposed.The electronic tongue and electronic nose were used to collect the fingerprint information of taste and smell for five different kinds of green tea.The one-dimensional electronic tongue and electronic nose signals were transformed into two-dimensional time-frequency maps using the short-time Fourier transform(STFT),which fully revealed the distribution characteristics of the signal energy in the time-frequency domain.A CNN-Transformer combination model was proposed to realize the fusion of the electronic tongue and the electronic nose information and pattern recognition.The model adopted selective kernel convolution and normalized attention in designing convolution module to replace the convolution layer of the traditional CNN to achieve the dynamic extraction of local features from the time-frequency map of the signal.The multi-head self-attention mechanism in the Transformer encoder was used to extract the global temporal information in the features of the electronic tongue and the electronic nose and achieve the weighted fusion of their features.Finally,classification recognition was carried out by the fully connected layer.[Results]The information fusion method based on electronic tongue and electronic nose could effectively extract the deep features of the taste and smell signals from green tea samples and provide richer fused feature representations for the model to achieve highly accurate recognition of different species,with a test set accuracy,precision,recall and F 1-Score of 99.00%,99.05%,99.00%,and 99.00%,respectively.[Conclusion]This study provides a low-cost,fast and efficient detection method for green tea species recognition.